Multi-level Fusion of FDG PET and MRI for Automated Epileptic Lesion Detection

Accurate localization of the epileptic lesion is crucial for surgical management of patients with epilepsy. However, detection rate of epileptic lesion with routine MRI imaging is limited. This study aims to investigate whether image-level and feature-level fusion of [¹⁸F]FDG PET and MR images using radiomics features would improve the detection of epileptic lesion. Forty-six drug refractory epilepsy patients with temporal and extra-temporal lesions who had PET/MRI exams for pre-surgical evaluation and follow-up MRI scans for post-surgical evaluation were included in this study, as well as 33 healthy controls who had PET/MRI exams. Radiomics features were extracted from high-resolution MRI, FDG PET, and fused images separately and combined. Image-level and feature-level fusions were applied to systematically search for optimal feature combination, which were then fed into the logistic regression models for performance evaluation. Models based on features extracted from fused images using Discrete Wavelet Transform showed better performance (top AUC = 0.871) with smaller feature counts, compared with those based on FDG PET alone (AUC = 0.838) or MRI alone (AUC = 0.763). Concatenated features using fused images together with original modalities further improved model performance, with the best model reaching AUC of 0.908.Our study suggests that multi-level fusion of FDG PET and T1w-MRI with radiomics feature extraction holds great potential in automated epileptic lesion detection with much higher performance over single modalities.Clinical Relevance— This study reveals that multi-level fusion of FDG PET and MRI with radiomics has potential to achieve automated epileptic lesion detection with much higher accuracy, potentially improving surgical outcomes.